<p>Scene text recognition (STR) is a computer vision task that predicts the text within wild images. Although a large amount of labeled text images is required to understand entire character sequences, acquiring such labeled data in real-world scenarios is challenging because of the high annotation cost. Therefore, previous studies have actively integrated STR with self-supervised learning, such as contrastive learning or generative modeling. However, they have limitations for STR. For example, contrastive learning can be easily affected by false negatives from frequent characters, and generative modeling may capture unnecessary features, such as background and noise. Additionally, both methods are unable to consider sequential features in text. To overcome these limitations, we propose a self-supervised learning framework for STR based on Information Maximization (STRIM). STRIM effectively learns both text and sequence information by reducing redundancy at both global and local levels. Notably, STRIM optimizes the sequential feature space with positional encoding, thereby capturing the sequential structure of text. Experimental results show that STRIM achieves competitive performance on benchmark datasets and demonstrates superior performance when labeled data is extremely limited.</p>